Hardware best practices

Talk about Hardware used for Deep Learning


User avatar
bryanlyon
Site Admin
Posts: 793
Joined: Fri Jul 12, 2019 12:49 am
Answers: 44
Location: San Francisco
Has thanked: 4 times
Been thanked: 218 times
Contact:

Re: RTX 3060

Post by bryanlyon »

stowaway wrote: Thu Oct 28, 2021 3:45 pm

I'm not having any luck with my Raedon (Asrock Radeon RX 6700 XT Challenger Pro 12G OC)

So I've decided I'm going to purchase a new video card.

Is the RTX 3060 chipset have fully compatiable?

Or can you suggest the best (mid range) card that can be purchased new (29/10/2021)

Unfortunately, due to current lack of availability getting ANY GPU is hard right now. But yes, any 20xx or 30xx GPU is acceptable and supported, and if you can get any in your preferred budget go for it.

User avatar
frankr69
Posts: 1
Joined: Tue May 12, 2020 9:08 am

New RTX 3050

Post by frankr69 »

Hi, is it productive to buy a new RTX 3050 8gb?

User avatar
bryanlyon
Site Admin
Posts: 793
Joined: Fri Jul 12, 2019 12:49 am
Answers: 44
Location: San Francisco
Has thanked: 4 times
Been thanked: 218 times
Contact:

Re: New RTX 3050

Post by bryanlyon »

frankr69 wrote: Thu Jan 27, 2022 11:41 am

Hi, is it productive to buy a new RTX 3050 8gb?

Should be fine (if you can get ahold of one).

User avatar
anselmBe
Posts: 9
Joined: Tue Feb 08, 2022 10:50 pm
Has thanked: 2 times

Re: Hardware best practices

Post by anselmBe »

Hurray!
I started my first Training Session on my beloved 1070 with 8Gigs.
But this is my Main Computer ... So having the Trainer running 24/7 for some upcoming time ahead. And after having MS restarting my PC because of Updates twice already, I started looking for Alternatives.
I do have an Oculus Rift certified ASUS PC with a 1060 and (only) 3 Gigs of RAM on the GPU.
My training Set wont start with the same settings.
So, I just orderd myself a 3060 with 8 Gigs:
MSI GeForce RTX 3060 Ti Gaming X 8G LHR @ 760€
Looking forward into Buliding a PC just for this FaceSwapping Pleasure.
I am thinking of using a HP Z440 as base Model. Stuff two SSDs and fill her up with RAM.
Running Ubuntu on it, so Windows won't keep me restarting.
Still not sure if the 3060 will run on the Z440 ... will see.
Cheers Guys :)

User avatar
anselmBe
Posts: 9
Joined: Tue Feb 08, 2022 10:50 pm
Has thanked: 2 times

Re: Hardware best practices

Post by anselmBe »

(Pardon my Language!)
FRACK ME!!!!!
Okay ...I just upgraded my PC from that (old) 1070 to the new 3060 TI.
I had run the old Training onthe 1070 for almost a week and had not one iterations

I am running now on the 3060 for (chekc the clock) three minutes and I have 450 Iterartions already!!!

Everybody that might thinking of taking this serious: invest in a new GPU. THe difference is like coming from the stoneage

User avatar
anselmBe
Posts: 9
Joined: Tue Feb 08, 2022 10:50 pm
Has thanked: 2 times

Re: Hardware best practices

Post by anselmBe »

It is 6 Minutes now and I start to see faces ...

User avatar
anselmBe
Posts: 9
Joined: Tue Feb 08, 2022 10:50 pm
Has thanked: 2 times

Re: Hardware best practices

Post by anselmBe »

Just one last milestone: 10 Minutes in: 1450 Iterations

User avatar
jcarl
Posts: 13
Joined: Sun Nov 13, 2022 6:41 am
Has thanked: 4 times
Been thanked: 1 time

Re: Hardware best practices

Post by jcarl »

A few GPU questions:

  1. Are the new 40xx GPUs automatically supported by Faceswap, or is an update required?

  2. Do faceswap workloads cause concerning memory temperatures on the 3090 (as they do with crypto mining)?

  3. How would the 4080 16GB would stack up against the 3090 or 3090ti?

User avatar
MaxHunter
Posts: 193
Joined: Thu May 26, 2022 6:02 am
Has thanked: 177 times
Been thanked: 13 times

Re: Hardware best practices

Post by MaxHunter »

As for the 4080:

I have a 3080ti and I just bought a 3090. The 4080 only has 4gbs more than a 3080ti - you are not going to get much more out of a 4080 than you would out of a 3080ti. The 3090 on the other hand is going to be twice as good for faceswapping and machine learning in my humble opinion because of it's higher memory. It can work more complex models, higher batch sizes, etc.

The 4080 is going to be better at gaming but not necessarily better at machine learning. Machine learning as I've come to learn relies more on memory and tensors than speed (and worse when you overclock.) So go for a 3090 if you're aim is machine learning - it's more bang for your buck. I picked up a 3090 on eBay for a little more than $800, where as a 4090 is over $2,000, for similar (slightly better) machine learning applications, or a 4080 which are worse than a 3090 and are going for almost twice as much.

User avatar
jcarl
Posts: 13
Joined: Sun Nov 13, 2022 6:41 am
Has thanked: 4 times
Been thanked: 1 time

Re: Hardware best practices

Post by jcarl »

Thank you for the reply. How are the RAM temps on your 3090? I read that VRAM intensive applications (e.g., ETH mining) caused dangerous temperature (>100 C memory junction temperature) levels on many 3090s, and required additional cooling to the back of the card. I wonder if FS has a similar type of memory intensive, VRAM heat inducing workload.

I have been looking at 3090tis over base 3090s mostly over possible temperature concerns.

User avatar
bryanlyon
Site Admin
Posts: 793
Joined: Fri Jul 12, 2019 12:49 am
Answers: 44
Location: San Francisco
Has thanked: 4 times
Been thanked: 218 times
Contact:

Re: Hardware best practices

Post by bryanlyon »

jcarl wrote: Sat Dec 10, 2022 2:41 am

Thank you for the reply. How are the RAM temps on your 3090? I read that VRAM intensive applications (e.g., ETH mining) caused dangerous temperature (>100 C memory junction temperature) levels on many 3090s, and required additional cooling to the back of the card. I wonder if FS has a similar type of memory intensive, VRAM heat inducing workload.

I have been looking at 3090tis over base 3090s mostly over possible temperature concerns.

They're fine. In mining you're doing more memory reads/weird than anything else. You're replacing everything in RAM multiple times per second.

In AI tasks you're never writing THAT hard and the ram will generally stay cool..

Just remember that these cards were designed (at least partially) for AI tasks and so it won't be outside the expected use case (unlike mining which is definitely more stress than they're designed for.)

That said, I'm not sure I'd recommend the 3090 or 3090 ti unless you have a VERY good reason to need that much ram.

User avatar
MaxHunter
Posts: 193
Joined: Thu May 26, 2022 6:02 am
Has thanked: 177 times
Been thanked: 13 times

Re: Hardware best practices

Post by MaxHunter »

To be honest, I haven't had a chance to test it... In fact the Post Office lost it today!!!!!!! 🤦🤯😫😣😖😤

But I'll let you know as soon I get a chance.

(Apparently I cursed myself by posting. 😆)

User avatar
jcarl
Posts: 13
Joined: Sun Nov 13, 2022 6:41 am
Has thanked: 4 times
Been thanked: 1 time

Re: Hardware best practices

Post by jcarl »

bryanlyon wrote: Sat Dec 10, 2022 4:03 am

That said, I'm not sure I'd recommend the 3090 or 3090 ti unless you have a VERY good reason to need that much ram.

GPU prices are still fluctuating quite a bit, but in my area, the 3090ti is hovering around the same price as the 4080 16GB. Would you recommend the 4080 over 3090ti?

User avatar
bryanlyon
Site Admin
Posts: 793
Joined: Fri Jul 12, 2019 12:49 am
Answers: 44
Location: San Francisco
Has thanked: 4 times
Been thanked: 218 times
Contact:

Re: Hardware best practices

Post by bryanlyon »

No. I tend to stick with recommending the 60 level of the GPUs. High end GPUs cost a LOT more for a BIT of better speed. Generally the 60-70 offers the best value.

If you just want a top-end GPU then by all means get a 3090 ti over a 4080. The 4080 is a new card and it usually takes a while to get good software support. (though Nvidia says that this generation is different... We'll see)

User avatar
doggydaddy
Posts: 6
Joined: Wed Jan 18, 2023 2:45 am
Has thanked: 3 times

Re: Hardware best practices

Post by doggydaddy »

I am considering the 3060 or the 3060 Ti to potentially replace my 2060...

  • the 3060Ti is about $100 more than the 3060

  • the 3060Ti only has 8GB VRAM while the 3060 has 12GB

  • the 3060Ti has significantly (perhaps 33%) more CUDA and Tensor cores than the 3060

which one is better for ML/faceswap? If I were really hardcore on gaming I would choose the Ti, but I dont know if that translates to faceswap

and i guess more generally, is the 3060 noticeably better for ML than the 2060? are there any benchmark tests?

thank you for your insight

User avatar
torzdf
Posts: 2649
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 128 times
Been thanked: 623 times

Re: Hardware best practices

Post by torzdf »

It depends what you want, the ability to train faster (in which case go for the Ti) or the ability to be able to train larger models (albeit slower than the Ti) then go for the 3060.

Both would be an upgrade on the 2060, yes.

My word is final

User avatar
trippod
Posts: 3
Joined: Mon Nov 13, 2023 8:23 am
Has thanked: 4 times

Re: Hardware best practices

Post by trippod »

hi, i want to start with faceswap profesionally, i want to buy a gpu this black friday, what about 4060 ti? worth it? or 3060 12 gb will be a best value? 16 gb with 2678v3 12c/24t is enough? thx.

User avatar
torzdf
Posts: 2649
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 128 times
Been thanked: 623 times

Re: Hardware best practices

Post by torzdf »

People's views will vary on this, but I would always favour VRAM over Cuda Cores. Whilst more Cuda Cores will train faster, more VRAM will mean being able to train larger models (at a time cost, obviously). Personally, I would go with the GPU that gives me the most VRAM for my buck.

My word is final

User avatar
trippod
Posts: 3
Joined: Mon Nov 13, 2023 8:23 am
Has thanked: 4 times

Re: Hardware best practices

Post by trippod »

then 16gb 4060ti could be a good deal?

User avatar
bryanlyon
Site Admin
Posts: 793
Joined: Fri Jul 12, 2019 12:49 am
Answers: 44
Location: San Francisco
Has thanked: 4 times
Been thanked: 218 times
Contact:

Re: Hardware best practices

Post by bryanlyon »

The trouble with hardware recommendations since the 20xx series has been that you always have to add "if you can get it for a good price" to the end of ANY recommendation. Will a 4060 TI with 16gb of vram work? Definitely. Will it beat a $100 1080? Yes. Is it worthwhile to buy it for FaceSwap? That's ENTIRELY up to you.

16gb of Vram is good enough for any of our models. Most of which were designed to work with 8gb 2060 tis or 1070s. The 16gb of ram will be plenty for most FaceSwap tasks.

Is the 4060ti fast enough? Unquestionably. I started training with 2x 970s in SLI. They'd take days to do what a 4060ti could do in hours.

Is it the right price? That's up to you. For some people spending ANY money on FaceSwap is spurious as it's just "for the memes". Others spend thousands on A6000s and consider them a bargain.

So I repeat: Will a 4060 TI with 16gb of vram work? Definitely. Will it beat a $100 1080? Yes. Is it worthwhile to buy it for FaceSwap? That's ENTIRELY up to you.

Post Reply